Open-ended discovery systems are not truly open-ended; CORAL is the first framework to make them autonomous

LLM (Large Language Model)-based evolution promises sustained, open-ended discovery: systems that search, accumulate knowledge, and improve indefinitely without human steering. Every existing framework quietly breaks that promise. Fixed heuristics and hard-coded exploration rules sit underneath the surface, pre-deciding how agents branch, when they pivot, and what they prioritize. The autonomy is cosmetic.

CORAL replaces that scaffolding with long-running agents operating through shared persistent memory and asynchronous multi-agent execution. This design shift has mechanical significance. Instead of a central controller dispatching tasks on a fixed schedule, agents explore and reflect independently, writing findings into shared memory that other agents can build on. Coordination happens through heartbeat-based interventions, health checks that detect stalled or looping agents and trigger recovery, rather than predetermined branching logic. Isolated workspaces and evaluator separation prevent agents from gaming their own fitness signals, a failure mode that collapses most self-improving systems. The framework targets mathematical, algorithmic, and systems-level problems, domains where progress requires sustained multi-step search rather than single-query retrieval.

The abstract truncates before reporting benchmark numbers, so performance claims against prior evolutionary LLM methods cannot be verified here. The structural novelty resides in the combined infrastructure layer of persistent memory, asynchronous execution, and health management, not in a single algorithmic idea. For teams building research automation or algorithm search pipelines, the practical gap CORAL closes is the agent collapse problem: systems that run for hours, drift into loops, and produce nothing recoverable. Heartbeat-based session management is a low-glamour fix that matters enormously in long-horizon deployments.

Key takeaways:

• Shared persistent memory plus asynchronous execution replaces fixed heuristics; agents accumulate and build on each other's findings rather than exploring in isolated, pre-scripted lanes.
• Hard-coded control flow is the hidden ceiling in existing open-ended discovery systems — removing it is a prerequisite for genuine sustained search, and not merely an architectural preference.
• Teams running long-horizon agent pipelines for algorithm search or automated research should audit whether their systems have evaluator separation and stall-recovery mechanisms before scaling compute.

Source: CORAL: Towards Autonomous Multi-Agent Evolution for Open-Ended Discovery